216 H. Cossette et al. Insurance: Mathematics and Economics 26 2000 213–222
Table 1 Extremal distributions in 3.1, two moments known, infinite spectrum
Value of s M
µ,σ
s W
µ,σ
s − M
µ,σ
s 0 s µ
σ
2
s − µ
2
+ σ
2
µ s δ
2
µ s
− µs µs
s δ
2
µ s
− µ
2
s − µ
2
+ σ
2
σ
2
s − µ
2
+ σ
2
positively orthant dependent POD, in short. For more details about POD, the interested reader is referred e.g. to Szekli 1995, pp. 144–145.
3. Stochastic bounds on a single loss: unknown marginals
We now examine the construction of bounds on F
X
when only the support and the first few moments of the marginals F
X
j
, j = 1, 2, . . . , m, are known. We first derive the best stochastic upper and lower bounds on the
X
j i
’s. For that purpose, we use the following results of Kaas and Goovaerts 1985. Consider a non-negative random variable Y for which only the mean µ, and the standard deviation σ are known.
Then, there exists two cumulative distribution functions, M
µ,σ
and W
µ,σ
, say, such that M
µ,σ
s ≤ F
Y
s ≤ W
µ,σ
s 3.1
is verified for all s ≥ 0. Explicit expressions for the extremal distributions in 3.1 are provided in Table 1 Table 2
in Kaas and Goovaerts, 1985, where δ
2
stands for the second moment of Y , i.e. δ
2
= EY
2
. When it is further known that there exists an upper bound b for Y , i.e. P [Y
≤ b] = 1, the extremal distributions in 3.1 can be refined as
M
µ,σ,b
s ≤ F
Y
s ≤ W
µ,σ,b
s, 3.2
which is verified for all s ≥ 0. Explicit expressions for these extremal distributions are provided in Table 2 Table
1 in Kaas and Goovaerts, 1985. When the skewness γ of Y is also known, tighter bounds M
µ,σ,γ
and W
µ,σ,γ
, say, can be found such that M
µ,σ,γ
s ≤ F
Y
s ≤ W
µ,σ,γ
s 3.3
is verified for all s ≥ 0. Explicit expressions for the extremal distributions in 3.3 are provided in Table 3 Table 4 in
Kaas and Goovaerts, 1985, where the following symbols are used: δ
3
stands for third moment of Y , i.e. δ
3
= EY
3
, β
1
s =
γ + 3σ
2
+ µ
3
− sδ
2
δ
2
− sµ ,
β
2
s =
δ
2
− µs µ
− s ,
and α
±
= δ
3
− µδ
2
± q
δ
3
− µδ
2 2
− 4σ
2
µδ
3
− δ
2 2
2σ
2
.
Table 2 Extremal distributions in 3.2, two moments known, finite spectrum
Value of s M
µ,σ,b
s W
µ,σ,b
s − M
µ,σ,b
s 0 s µ
− σ
2
b − µ
σ
2
s − µ
2
+ σ
2
µ − σ
2
b − µ s δ
2
µ σ
2
+ µ − bµ − ssb σ
2
+ µµ − bss − b s δ
2
µ s
− µ
2
s − µ
2
+ σ
2
σ
2
s − µ
2
+ σ
2
H. Cossette et al. Insurance: Mathematics and Economics 26 2000 213–222 217
Table 3 Extremal distributions in 3.3, three moments known, infinite spectrum
Value of s M
µ,σ,γ
s W
µ,σ,γ
s − M
µ,σ,γ
s 0 s α
−
µ − β
2
ss − β
2
s α
−
s δ
2
µ σ
2
+ µ − sµ − β
1
ssβ
1
s σ
2
+ µ − β
1
sµss − β
1
s δ
2
µ s α
+
µ − sβ
2
s − s
µ − β
2
ss − β
2
s s α
+
σ
2
+ µ − sµ − β
1
ssβ
1
s + σ
2
+ µ − sµβ
1
s − sβ
1
s σ
2
+ µ − β
1
sµss − β
1
s
As in 3.2, when it is further known that there exists an upper bound b for Y , the extremal distributions in 3.3 can be refined as
M
µ,σ,γ ,b
s ≤ F
Y
s ≤ W
µ,σ,γ ,b
s, 3.4
which is verified for all s ≥ 0. Explicit expressions for these extremal distributions are provided in Table 4 Table
3 in Kaas and Goovaerts, 1985, where the following symbols are used: ζ
= γ
+ 3σ
2
+ µ
3
− bδ
2
δ
2
− bµ ,
β
∗ 2
s =
γ + 3σ
2
+ µ
3
− b + sδ
2
+ bsµ δ
2
− b + sµ + bs .
Now, let µ
j
, σ
j
and γ
j
be the mean, the standard deviation and the skewness corresponding to F
X
j
, j
= 1, 2, . . . , m. With Tables 1 and 3, we can find the best upper and lower bounds on F
X
j
, j = 1, 2, . . . , m,
namely M
j
s ≤ F
X
j
s ≤ W
j
s for j
= 1, 2, . . . , m, 3.5
where M
j
resp. W
j
stands for either M
µ
j
,σ
j
or M
µ
j
,σ
j
,γ
j
resp. W
µ
j
,σ
j
or W
µ
j
,σ
j
,γ
j
. When there exist b
j
such that F
X
j
b
j
= 1, j = 1, 2, . . . , m, then 3.5 becomes M
b
j
j
s ≤ F
X
j
s ≤ W
b
j
j
s for j
= 1, . . . , m, 3.6
where M
b
j
j
resp. W
b
j
j
stands for
either M
µ
j
,σ
j
,b
j
or M
µ
j
,σ
j
,γ
j
,b
j
resp. W
µ
j
,σ
j
,b
j
resp. W
µ
j
,σ
j
,γ
j
,b
j
, j = 1, 2, . . . , m. Then,
˜ F
min
s ≤ F
X
s ≤ ˜
F
max
s for all s
≥ 0, 3.7
with, for s ∈ R,
˜ F
min
s =
sup
x
1
,x
2
,... ,x
m
∈ Ps
max
m
X
j =1
lim
n →∞
M
j
x
j
− 1
n − m − 1, 0
,
Table 4 Extremal distributions in 3.4, three moments known, finite spectrum
Value of s M
µ,σ,γ ,b
s W
µ,σ,γ ,b
s − M
µ,σ,γ ,b
s 0 s α
−
σ
2
+ µ − β
∗ 2
sµ − bs − β
∗ 2
ss − b
α
−
s ζ σ
2
+ µ − sµ − β
1
ssβ
1
s σ
2
+ µ − β
1
sµss − β
1
s ζ s α
+
σ
2
+ µ − sµ − bβ
∗ 2
s − sβ
∗ 2
s − b
σ
2
+ µ − β
∗ 2
sµ − bs − β
∗ 2
ss − b
s α
+
σ
2
+ µ − sµ − β
1
ssβ
1
s + σ
2
+ µ − sµβ
1
s − sβ
1
s σ
2
+ µ − β
1
sµss − β
1
s
218 H. Cossette et al. Insurance: Mathematics and Economics 26 2000 213–222
and ˜
F
max
s =
inf
x
1
,x
2
,... ,x
m
∈ Ps
min
m
X
j =1
W
j
x
j
, 1
. Of course, when there exist b
j
such that F
X
j
b
j
= 1, j = 1, . . . , m, M
j
and W
j
in the expressions of ˜ F
min
and ˜
F
max
are replaced by M
b
j
j
and W
b
j
j
, respectively. If the X
1 i
, X
2 i
, . . . , X
m i
’s are POD, then 3.7 is valid with the following improved bounds:
˜ F
min
s =
sup
x
1
,x
2
,... ,x
m
∈ Ps
m
Y
j =1
M
j
x
j
and ˜
F
max
s = 1 −
sup
x
1
,x
2
,... ,x
m
∈ Ps
m
Y
j =1
1 − W
j
x
j
.
4. Stochastic bounds on the total amount of loss